Ando–Hiai Type Inequality

• Ando–Hiai type inequality is a fundamental result in operator theory that defines order comparisons for powered operator means through necessary and sufficient conditions.
• It employs the Kubo–Ando framework and power means to systematically extend operator inequalities to multivariate, perspective, and quantum trace settings.
• Concrete examples such as logarithmic, identric, and Heinz means illustrate its broad applicability and its role in deriving norm and entropy inequalities.

The Ando-Hiai type inequality is a central structural result in operator theory, particularly in the context of matrix and operator means defined via the Kubo–Ando framework. It provides order-comparisons of powered operator means and enables systematic characterization and extension to multivariate and perspective settings. Fundamental to its utility are precise necessary and sufficient conditions for equality, a rich algorithmic representation via power means, and deep connections to operator monotonicity and quantum trace inequalities.

1. Operator Means: Kubo–Ando Theory and Integral Representation

Operator means are binary maps $$\sigma: \mathrm{PS} \times \mathrm{PS} \to \mathrm{PS}$$ on the cone $\mathrm{PS}$ of positive semi-definite operators on a Hilbert space $H$, adhering to monotonicity, transformer-inequality, upper semi-continuity, and normalization $I \sigma I = I$. Kubo and Ando established that each $\sigma$ is uniquely specified by an operator-monotone function $f$ with $f(1)=1$ via

$$\sigma(A, B) = A^{1/2} f(A^{-1/2}BA^{-1/2}) A^{1/2}.$$

Yamazaki further proved every operator mean has an integral representation: $f(x) = \int_0^1 p_t(\lambda; x) \, d\mu(\lambda),$ where

$$p_t(\lambda;x) = \left[(1-\lambda) + \lambda x^t\right]^{1/t}$$

and $\mu$ is a Borel probability measure. This formula interpolates arithmetic, geometric, and harmonic means and extends in the limit $t\to0$ to

$f(x) = \int_0^1 x^\lambda d\mu(\lambda).$

Thus, $\sigma(A,B)$ admits an operator-valued integral via power means (Yamazaki, 2018).

2. Statement and Generalization of the Ando–Hiai Inequality

Classically, the Ando–Hiai inequality asserts for $A, B \in \mathrm{PS}$ and any $\lambda \in [0,1]$

$$A \#_{\lambda} B \leq I \implies A^r \#_{\lambda} B^r \leq I \quad \forall r \geq 1,$$

where $A \#_{\lambda} B$ denotes the $\lambda$-weighted geometric mean. For general operator means $\sigma$,

$$\sigma(A,B) \leq I \implies \sigma(A^r, B^r) \leq I \quad \forall r \geq 1.$$

Yamazaki characterized exactly which operator means possess the Ando–Hiai property in terms of the representing function $f$: $\sigma$ has the Ando–Hiai property if and only if

$x f'(1) \leq f(x), \quad \forall x > 0,$

which further equates to the differential-inequality $f(x)^r \leq f(x^r)$, linking function, operator, and integral representations (Yamazaki, 2018).

3. Proof Structure and Characterization

Yamazaki’s argument proceeds via:

• Compact convexity of $C_t = \{f \in M : p_t(f'(1);x) \leq f(x)\}$; the extreme points are power-mean functions.
• The Kreĭn–Milman theorem guarantees that any $f \in C_t$ admits the power-mean integral representation.
• For fixed $r \geq 1$, using convexity $s \mapsto s^r$ yields $p_t(\lambda;x)^r \leq p_t(\lambda;x^r)$, and integrating gives $f(x)^r \leq f(x^r)$.
• The key lemma: $f(x)^r \leq f(x^r)$ for all $x > 0$, $r \geq 1$ is equivalent to $x f'(1) \leq f(x)$, ensuring the operator inequality property (Yamazaki, 2018).

4. Examples and Explicit Operator Means

The criterion is satisfied for several canonical means:

Mean	Representing Function $f(x)$	$x f'(1) \leq f(x)$ verified
Logarithmic Mean	$(x-1)/\ln x$	Yes
Identric Mean	$\exp[(x\ln x)/(x-1) - 1]$	Yes
Heinz Mean	$[x^t + x^{1-t}]/2$	Yes for $t \in [0,1]$

Each satisfies $$\sigma(A,B) \leq I \implies \sigma(A^r, B^r) \leq I$$ for all $r \geq 1$ (Yamazaki, 2018).

5. Variants, Perspectives, and Extensions

Extensions to multivariate and deformed operator means incorporate similar properties. Power means, Karcher means, and perspectives admit both direct and complementary Ando–Hiai-type inequalities given power-monotonicity (pmi) of their representing function: $f(x^p) \geq f(x)^p, \quad p \geq 1.$ Furthermore, operator perspectives $P_f(A,B) = B^{1/2} f(B^{-1/2}AB^{-1/2}) B^{1/2}$ inherit the property when $f$ is pmi, yielding

$$P_f(A,B)\leq I \implies P_f(A^p,B^p)\leq I, \quad p \in (0,1].$$

Refined Ando–Hiai inequalities for non-invertible operators, deformed means, and matrix functions such as log-Euclidean means have been established in (Hiai et al., 2019, Kian et al., 2019, Hiai et al., 2018), and other works.

6. Spectral Geometric Means and Restricted Ando–Hiai Property

Recent research treats the spectral geometric mean $A \natural_t B$ and related two-variable operator functions $F_{k,t}(A,B)$, showing Ando–Hiai type inequalities hold only for $q$ bounded by explicit functions of the parameters $(k,t)$: $$F_{k,t}(A,B) \leq I \implies F_{k,t}(A^q,B^q) \leq I, \quad 0 < q \leq \frac{2ktL}{1-2kt} \leq 1,\, L = 1+2t-4kt,$$ with analogous bounds for $A \natural_t B$ (Seo et al., 28 Dec 2025). These restricted ranges are shown to be sharp, with counterexamples outside the stated interval.

7. Connections to Operator Monotone Functions and Trace Inequalities

The Ando–Hiai paradigm is tightly interwoven with operator monotonicity. The converse established by Ando and Hiai is that if $f(A_+ B) \ge f(A \sigma B)$ for a symmetric mean $\sigma \neq A_+$ (or $f(A_- B) \le f(A \sigma B)$ for $\sigma \neq A_-$), then $f$ must be operator monotone (Dinh et al., 2018). This criterion is reinforced by geometric mean analogues, self-adjoint means, and chain inequalities involving Heron and Heinz means. Variants extend this approach to trace inequalities, log-majorizations, and norm inequalities, with further generalizations to Golden–Thompson type bounds (Carlen et al., 2022, Hiai, 2015).

• Yamazaki, "An integral representation of operator means via the power means and an application to the Ando–Hiai inequality" (Yamazaki, 2018)
• Hiai, Seo, Wada, "Ando–Hiai type inequalities for operator means and operator perspectives" (Hiai et al., 2019)
• Moradi, Furuichi, Sababheh, "Operator Spectral Geometric Versus Geometric Mean" (Moradi et al., 2021)
• Seo, Wada, Yamazaki, "On the Ando–Hiai property for spectral geometric means" (Seo et al., 28 Dec 2025)
• Kian, Moslehian, "Power means of probability measures and Ando–Hiai inequality" (Kian et al., 2018)

The Ando–Hiai type inequality constitutes a cornerstone for quantitative matrix analysis, operator means, and quantum information theory, providing both a deep conceptual understanding and practical methods for extending operator monotonicity, norm bounds, and entropy inequalities.